前一篇完成了 TimesFM 的运行
TimesFM（Time Series Foundation Model）安装（2）-CSDN博客文章浏览阅读520次，点赞13次，收藏24次。决定在 小红帽ubuntu UBUNTU安装 timesFM在 ide.cloud.tencent.com 的环境上进行安装 环境 慎选环境，确保>16G安装Conda 3.10 python重要步骤 安装 pyenv and poetry确认已经完成安装这里安装完需要设置环境变量，如果不能看到 version 版本时Add `export PATH="/root/.local/bin:$PATH"` to your shell configuratiohttps://blog.csdn.net/chenchihwen/article/details/144386472?sharetype=blogdetail&sharerId=144386472&sharerefer=PC&sharesource=chenchihwen&spm=1011.2480.3001.8118
这篇针对里面运行的数据进行分析
根据代码 数据是来自 exp, exp 是来自引用 的 .utils.py
这是 utilis.py 的代码
是从Nixtla来的 尼克斯塔与时间序列预测
"""Forked from https://github.com/Nixtla/nixtla/blob/main/experiments/amazon-chronos/src/utils.py."""
# Copyright 2024 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License."""Forked from https://github.com/Nixtla/nixtla/blob/main/experiments/amazon-chronos/src/utils.py."""from functools import partial
from itertools import repeat
import multiprocessing
import os
from pathlib import Path
from typing import Listfrom gluonts.dataset import Dataset
from gluonts.dataset.repository.datasets import (dataset_names as gluonts_datasets,get_dataset,
)
from gluonts.time_feature.seasonality import get_seasonality
import numpy as np
import pandas as pd
from utilsforecast.evaluation import evaluate
from utilsforecast.losses import mae, mase, smapedef parallel_transform(inp):ts, last_n = inp[0], inp[1]return ExperimentHandler._transform_gluonts_instance_to_df(ts, last_n=last_n)def quantile_loss(df: pd.DataFrame,models: list,q: float = 0.5,id_col: str = "unique_id",target_col: str = "y",
) -> pd.DataFrame:delta_y = df[models].sub(df[target_col], axis=0)res = (np.maximum(q * delta_y, (q - 1) * delta_y).groupby(df[id_col], observed=True).mean())res.index.name = id_colres = res.reset_index()return resclass ExperimentHandler:def __init__(self,dataset: str,quantiles: List[float] = list(np.arange(1, 10) / 10.0),results_dir: str = "./results",models_dir: str = "./models",):if dataset not in gluonts_datasets:raise Exception(f"dataset {dataset} not found in gluonts "f"available datasets: {', '.join(gluonts_datasets)}")self.dataset = datasetself.quantiles = quantilesself.level = self._transform_quantiles_to_levels(quantiles)self.results_dir = results_dirself.models_dir = models_dir# defining datasetsself._maybe_download_m3_or_m5_file(self.dataset)gluonts_dataset = get_dataset(self.dataset)self.horizon = gluonts_dataset.metadata.prediction_lengthif self.horizon is None:raise Exception(f"horizon not found for dataset {self.dataset} ""experiment cannot be run")self.freq = gluonts_dataset.metadata.freq# get_seasonality() returns 1 for freq='D', override this to 7. This significantly improves the accuracy of# statistical models on datasets like m5/nn5_daily. The models like AutoARIMA/AutoETS can still set# seasonality=1 internally on datasets like weather by choosing non-seasonal models during model selection.if self.freq == "D":self.seasonality = 7else:self.seasonality = get_seasonality(self.freq)self.gluonts_train_dataset = gluonts_dataset.trainself.gluonts_test_dataset = gluonts_dataset.testself._create_dir_if_not_exists(self.results_dir)try:multiprocessing.set_start_method("spawn")except RuntimeError:print("Multiprocessing context has already been set.")@staticmethoddef _maybe_download_m3_or_m5_file(dataset: str):if dataset[:2] == "m3":m3_file = Path.home() / ".gluonts" / "datasets" / "M3C.xls"if not m3_file.exists():from datasetsforecast.m3 import M3from datasetsforecast.utils import download_filedownload_file(m3_file.parent, M3.source_url)elif dataset == "m5":m5_raw_dir = Path.home() / ".gluonts" / "m5"if not m5_raw_dir.exists():import zipfilefrom datasetsforecast.m5 import M5from datasetsforecast.utils import download_filedownload_file(m5_raw_dir, M5.source_url)with zipfile.ZipFile(m5_raw_dir / "m5.zip", "r") as zip_ref:zip_ref.extractall(m5_raw_dir)@staticmethoddef _transform_quantiles_to_levels(quantiles: List[float]) -> List[int]:level = [int(100 - 200 * q) for q in quantiles if q < 0.5] # in this case mean=mediainlevel = sorted(list(set(level)))return level@staticmethoddef _create_dir_if_not_exists(directory: str):Path(directory).mkdir(parents=True, exist_ok=True)@staticmethoddef _transform_gluonts_instance_to_df(ts: dict,last_n: int | None = None,) -> pd.DataFrame:start_period = ts["start"]start_ds, freq = start_period.to_timestamp(), start_period.freqtarget = ts["target"]ds = pd.date_range(start=start_ds, freq=freq, periods=len(target))if last_n is not None:target = target[-last_n:]ds = ds[-last_n:]ts_df = pd.DataFrame({"unique_id": ts["item_id"], "ds": ds, "y": target})return ts_df@staticmethoddef _transform_gluonts_dataset_to_df(gluonts_dataset: Dataset,last_n: int | None = None,) -> pd.DataFrame:with multiprocessing.Pool(os.cpu_count()) as pool: # Create a process poolresults = pool.map(parallel_transform, zip(gluonts_dataset, repeat(last_n)))df = pd.concat(results)df = df.reset_index(drop=True)return df@propertydef train_df(self) -> pd.DataFrame:train_df = self._transform_gluonts_dataset_to_df(self.gluonts_train_dataset)return train_df@propertydef test_df(self) -> pd.DataFrame:test_df = self._transform_gluonts_dataset_to_df(self.gluonts_test_dataset,last_n=self.horizon,)# Make sure that only the first backtest window is used for evaluation on `traffic` / `exchange_rate` datasetsreturn test_df.groupby("unique_id", sort=False).head(self.horizon)def save_dataframe(self, df: pd.DataFrame, file_name: str):df.to_csv(f"{self.results_dir}/{file_name}", index=False)def save_results(self, fcst_df: pd.DataFrame, total_time: float, model_name: str):self.save_dataframe(fcst_df,f"{model_name}-{self.dataset}-fcst.csv",)time_df = pd.DataFrame({"time": [total_time], "model": model_name})self.save_dataframe(time_df,f"{model_name}-{self.dataset}-time.csv",)def fcst_from_level_to_quantiles(self,fcst_df: pd.DataFrame,model_name: str,) -> pd.DataFrame:fcst_df = fcst_df.copy()cols = ["unique_id", "ds", model_name]for q in self.quantiles:if q == 0.5:col = f"{model_name}"else:lv = int(100 - 200 * q)hi_or_lo = "lo" if lv > 0 else "hi"lv = abs(lv)col = f"{model_name}-{hi_or_lo}-{lv}"q_col = f"{model_name}-q-{q}"fcst_df[q_col] = fcst_df[col].valuescols.append(q_col)return fcst_df[cols]def evaluate_models(self, models: List[str]) -> pd.DataFrame:fcsts_df = []times_df = []for model in models:fcst_method_df = pd.read_csv(f"{self.results_dir}/{model}-{self.dataset}-fcst.csv").set_index(["unique_id", "ds"])fcsts_df.append(fcst_method_df)time_method_df = pd.read_csv(f"{self.results_dir}/{model}-{self.dataset}-time.csv")times_df.append(time_method_df)fcsts_df = pd.concat(fcsts_df, axis=1).reset_index()fcsts_df["ds"] = pd.to_datetime(fcsts_df["ds"])times_df = pd.concat(times_df)return self.evaluate_from_predictions(models=models, fcsts_df=fcsts_df, times_df=times_df)def evaluate_from_predictions(self, models: List[str], fcsts_df: pd.DataFrame, times_df: pd.DataFrame) -> pd.DataFrame:test_df = self.test_dftrain_df = self.train_dftest_df = test_df.merge(fcsts_df, how="left")assert test_df.isna().sum().sum() == 0, "merge contains nas"# point evaluationpoint_fcsts_cols = ["unique_id", "ds", "y"] + modelstest_df["unique_id"] = test_df["unique_id"].astype(str)train_df["unique_id"] = train_df["unique_id"].astype(str)mase_seas = partial(mase, seasonality=self.seasonality)eval_df = evaluate(test_df[point_fcsts_cols],train_df=train_df,metrics=[smape, mase_seas, mae],)# probabilistic evaluationeval_prob_df = []for q in self.quantiles:prob_cols = [f"{model}-q-{q}" for model in models]eval_q_df = quantile_loss(test_df, models=prob_cols, q=q)eval_q_df[prob_cols] = eval_q_df[prob_cols] * self.horizoneval_q_df = eval_q_df.rename(columns=dict(zip(prob_cols, models)))eval_q_df["metric"] = f"quantile-loss-{q}"eval_prob_df.append(eval_q_df)eval_prob_df = pd.concat(eval_prob_df)eval_prob_df = eval_prob_df.groupby("metric").sum().reset_index()total_y = test_df["y"].sum()eval_prob_df[models] = eval_prob_df[models] / total_yeval_prob_df["metric"] = "scaled_crps"eval_df = pd.concat([eval_df, eval_prob_df]).reset_index(drop=True)eval_df = eval_df.groupby("metric").mean(numeric_only=True).reset_index()eval_df = eval_df.melt(id_vars="metric", value_name="value", var_name="model")times_df.insert(0, "metric", "time")times_df = times_df.rename(columns={"time": "value"})eval_df = pd.concat([eval_df, times_df])eval_df.insert(0, "dataset", self.dataset)eval_df = eval_df.sort_values(["dataset", "metric", "model"])eval_df = eval_df.reset_index(drop=True)return eval_dfif __name__ == "__main__":multiprocessing.set_start_method("spawn")
以 toursim 月预测来分析
相关数据可以查看我上传的资源
TimesFM 预测数据来源 TimesFM（时间序列基础模型）是由谷歌研究开发的一种预训练时间序列基础模型
https://download.csdn.net/download/chenchihwen/90124776?spm=1001.2014.3001.5503这里已经将最原始的的 json 转换成 excel 格式
我们来看training 原始的 [data.json.gz] 内容如下：
通过 tourism_monthly 的 metadata.json
{"freq": "M", "target": null, "feat_static_cat": [{"name": "feat_static_cat_0", "cardinality": "366"}], "feat_static_real": [], "feat_dynamic_real": [], "feat_dynamic_cat": [], "prediction_length": 24}
代码转译后 excel 如下，
T1 标签最后的 training 资料 是1992/7/31
我们来看下预测的结果，与实际数据的比较，还行
| unique_id | ds | y | unique_id | ds | timesfm | ||
| T1 | 1992/8/31 | 6611.1 | T1 | 1992/8/31 | 5975.3 | ||
| T1 | 1992/9/30 | 4150.2 | T1 | 1992/9/30 | 4250.7 | ||
| T1 | 1992/10/31 | 2841.0 | T1 | 1992/10/31 | 2843.6 | ||
| T1 | 1992/11/30 | 1813.4 | T1 | 1992/11/30 | 2144.1 | ||
| T1 | 1992/12/31 | 2261.1 | T1 | 1992/12/31 | 2206.6 | ||
| T1 | 1993/1/31 | 1873.6 | T1 | 1993/1/31 | 1862.6 | ||
| T1 | 1993/2/28 | 1772.8 | T1 | 1993/2/28 | 1961.5 | ||
| T1 | 1993/3/31 | 2049.6 | T1 | 1993/3/31 | 2007.9 | ||
| T1 | 1993/4/30 | 2932.3 | T1 | 1993/4/30 | 2531.2 | ||
| T1 | 1993/5/31 | 3113.3 | T1 | 1993/5/31 | 2908.7 | ||
| T1 | 1993/6/30 | 3461.5 | T1 | 1993/6/30 | 3474.2 | ||
| T1 | 1993/7/31 | 6265.7 | T1 | 1993/7/31 | 5924.7 | ||
| T1 | 1993/8/31 | 6857.8 | T1 | 1993/8/31 | 6098.7 | ||
| T1 | 1993/9/30 | 4346.1 | T1 | 1993/9/30 | 4134.4 | ||
| T1 | 1993/10/31 | 3154.7 | T1 | 1993/10/31 | 2737.1 | ||
| T1 | 1993/11/30 | 2142.2 | T1 | 1993/11/30 | 1909.0 | ||
| T1 | 1993/12/31 | 2375.7 | T1 | 1993/12/31 | 2115.6 | ||
| T1 | 1994/1/31 | 1981.1 | T1 | 1994/1/31 | 1823.7 | ||
| T1 | 1994/2/28 | 1959.9 | T1 | 1994/2/28 | 1850.2 | ||
| T1 | 1994/3/31 | 2466.3 | T1 | 1994/3/31 | 1935.1 | ||
| T1 | 1994/4/30 | 2851.7 | T1 | 1994/4/30 | 2440.0 | ||
| T1 | 1994/5/31 | 3671.8 | T1 | 1994/5/31 | 2753.3 | ||
| T1 | 1994/6/30 | 3806.8 | T1 | 1994/6/30 | 3497.1 | ||
| T1 | 1994/7/31 | 6995.0 | T1 | 1994/7/31 | 5773.2 |
实际跑出的预计数是长得这个样子， 会有 timesfm-q-0.1 : timesfm-q-0.9
是因为代码中设置了
在Python代码中，尤其是涉及到时间序列分析、预测等相关场景（从你之前提供的看起来和时间序列相关的代码链接推测），`quantiles` 通常有以下一些常见用途：
### 概率预测与不确定性量化方面
1. **表示预测区间**：在很多预测任务中，模型给出的不只是一个单一的预测值（比如预测某个时间点的数值），而是一个预测区间来体现不确定性。`quantiles` 可以用来定义这些区间的边界，例如，常见的会使用 0.05、0.5、0.95 等不同分位数对应的数值来表示预测区间的下限、中位数、上限。像如果预测某商品未来销量，通过计算不同 `quantiles` 的值，能知道销量大概率（比如 90% 的置信区间，对应 0.05 和 0.95 分位数）会落在哪个范围，以及最有可能的中间值（0.5 分位数即中位数）是多少。
2. **评估预测的不确定性**：在评估预测模型好坏时，除了看预测值与实际值的偏差（比如均方误差等指标衡量准确性），还需要考量预测的不确定性是否合理。通过比较模型输出的 `quantiles` 对应的预测区间和实际观测值落在该区间的比例等情况，可以判断模型对不确定性的把握能力。例如，如果模型声称某个 `quantiles` 对应的区间有 90% 的概率包含真实值，但实际多次验证下来真实值只有 50% 的情况落在该区间，那就说明模型对不确定性的估计可能不太准确。
timesfm forecast 是取中位数 也就是 timesfm-q-0.5
